Research Pipeline

• You have a research direction or topic and want the system to autonomously discover viable ideas, implement experiments, and produce a reviewable paper draft — the "full pipeline" use case.
• You want to run multi-seed GPU experiments overnight and wake up to auto-reviewed results without manual intervention.
• You need an end-to-end reproducible research artifact: ranked ideas, experiment logs, review history, and LaTeX source all in one coherent output.
• You are targeting a specific venue (ICLR, NeurIPS, ICML, CVPR, ACL, etc.) and want the pipeline to write and compile a venue-formatted PDF automatically.

• You only need one stage in isolation (e.g., just idea discovery, just running experiments, or just paper writing) — invoke the individual skills (/idea-discovery, /run-experiment, /paper-writing) directly instead.
• You do not have GPU resources or remote servers configured, as Stages 3–4 require experiment deployment infrastructure.
• Your research requires manual figure creation before paper writing — the pipeline will pause and wait, so factor this into your timeline.

Orchestrates Workflow 1 (/idea-discovery), Workflow 2 (/run-experiment + /auto-review-loop), and optional Workflow 3 (/paper-writing) in sequence. Each workflow produces self-contained output files so you can inspect or resume at any stage.

AUTO_PROCEED=true auto-selects the top-ranked idea and continues without waiting; AUTO_PROCEED=false pauses at Gate 1 for explicit user confirmation. HUMAN_CHECKPOINT=true surfaces review scores after each round so you can provide custom fix instructions before the loop continues.

The auto-review loop supports three adversarial levels: medium (standard review), hard (reviewer memory + debate protocol), and nightmare (GPT reads the repo directly via codex exec + memory + debate). Up to 4 rounds of review/fix cycles run autonomously, targeting a score ≥ 6/10.

Small batches (≤5 jobs) are dispatched via /run-experiment; large multi-seed sweeps (≥10 jobs) are routed to /experiment-queue with OOM retry, stale-screen cleanup, phase dependencies, and crash-safe state. GPU availability is checked before deployment.

When AUTO_WRITE=true and VENUE is set, the pipeline automatically invokes /paper-writing after Stage 5, running plan → figure → write → compile → improvement loop phases and producing a final paper/main.pdf.

Every stage produces versioned, manifest-logged output files: IDEA_REPORT.md, AUTO_REVIEW.md, NARRATIVE_REPORT.md, and a full Research Pipeline Report summarizing GPU hours, review rounds, scores, and remaining TODOs.

Set AUTO_PROCEED=true, HUMAN_CHECKPOINT=false, and launch before bed. The pipeline discovers ideas, auto-selects the best, implements experiments, deploys to your GPU server, runs up to 4 review/fix rounds, and generates NARRATIVE_REPORT.md by morning.

Set AUTO_PROCEED=false to pause at Gate 1, review the ranked ideas yourself, pick or combine ideas, then approve. Stages 2–4 then run fully autonomously — you get human control over the research direction without babysitting the experiments.

Set AUTO_WRITE=true and VENUE=NeurIPS (or ICLR, ICML, CVPR, ACL, etc.). After experiments and review loops complete, the pipeline writes, compiles, and improves a venue-formatted LaTeX paper, ending with paper/main.pdf ready for submission.

Set REVIEWER_DIFFICULTY=nightmare to have GPT read the repository directly and apply memory + debate protocols during review rounds. Use this when you want the harshest possible pre-submission feedback before targeting a competitive venue.